uratori-ja-4b
A Japanese decision model from tokimoa. Qwen3.5-4B with a merged LoRA and a scoring head read at a marker after each option returns calibrated yes/no, choice and score distributions in one forward pass, built for grounding checks, document comparison and RAG judgments.
Japanese only. Needs trust_remote_code and about 12 GB of GPU memory. The maker's evaluation labels are majority votes of a draft answer and three LLM judges, not human labels. Weaker on relative dates, numbers and long inputs; trained once, on a single seed. Siblings: uratori-ja-310m (CPU, ONNX), 2b and 9b (0.887 on the same test split), plus MLX builds. On the same items the maker's own run of Jev 1.13 scores 0.903.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one probability
- classify — assigns a category from a fixed taxonomy
At a glance
| Parameters | 4.21B |
| Base model | Qwen/Qwen3.5-4B |
| Maker | tokimoa |
| Released | 2026-10-06 |
| License | apache-2.0 |
| Reported accuracy | 86.7% |
| Reported latency | about 100 ms per question on an RTX 5090 at batch 8 |
Get the weights
pip install systemonemodels
systemone pull tokimoa/uratori-ja
The files are served from the maker's Hugging Face repository, tokimoa/uratori-ja-4b, and verified against the checksums recorded here.